Assessment of multiple air pollutant exposures using a mobile monitoring approach: a comparison of spatial modeling with multiple regression algorithms

Abstract

Background/Aim: Geo-statistical models have been widely applied to assess fine-scale air pollution exposures in epidemiological studies, most of which were developed for criteria pollutants rather than other toxic air pollutants that may also be harmful for human health. Furthermore, the predictive ability of different algorithms in air pollution models have rarely been compared. We aim to develop short-term spatial models for 11 air exposure metrics and compare their performances using 6 modeling algorithms. Methods: We used a mobile laboratory with fast-response monitors to measure multiple gaseous pollutants (nitrogen dioxide, carbon monoxide, sulfur dioxide, ozone, benzene, toluene, methanol) and PM species (black carbon, surface area, and count- and volume-concentrations of ultrafine particles) in the Beijing metropolitan area. Data were collected from 130 repeated short-term monitoring locations for up to 31 days between July and September during the 2008 summer Olympics period and were calibrated with a central fixed site. We developed 6 spatial models for each pollutant using linear regression, dimension reduction, non-linear regression (NLR) and machine learning (ML) algorithms and included extensive geographic variables. Best models were selected using ten-fold cross-validation (CV). Results: The best models based on the largest CV R2 explained more than 60% of the variation for all the exposure metrics (range: 0.61 for methanol to 0.88 for ozone). Among the algorithms, random forest (RF, a ML algorithm) and the generalized additive model (GAM, a NLR algorithm) outperformed the other approaches for most of the pollutants (6 pollutants for RF and 3 for GAM). Incorporating a kriging model for model residuals improved the CV R2 by 3%~39%. Conclusions: Exposure models, especially based on ML and NLR algorithms, captured spatial variability of short-term average concentrations, had adequate predictive validity, and could be successfully applied to existing human health studies carried out during the Beijing Olympics period.